Pipeline parallelism is essential for training large neural networks, but existing schedules trade off throughput, memory, and optimization consistency. Synchronous pipelines preserve forward/backward weight consistency but suffer from bubbles; asynchronous pipelines remove bubbles but introduce weight-version mismatch, typically requiring weight stashing, prediction, or correction mechanisms. We introduce PACI (Pipeline Asynchronous training with Controlled Inconsistency), a bubble-free asynchronous pipeline method that bounds forward/backward version drift without weight stashing, prediction, additional parameter copies, or global synchronization. The key idea is to use local gradient accumulation as a version-control mechanism: by slowing parameter-version evolution relative to pipeline delay, PACI limits the number of optimizer updates crossed by any micro-batch while preserving steady-state utilization. In GPT-style language-model pretraining, PACI matches the stability and final perplexity of synchronous 1F1B-flush, retains the same peak memory footprint, achieves fully utilized pipeline throughput, and improves training time-to-accuracy by up to $1.69\times$ over the fastest flush baseline. These results show that forward/backward inconsistency need not be eliminated: when explicitly bounded, it can be safely traded for substantial efficiency gains.

The KV cache is the memory bottleneck of long-horizon language model deployment. Practically, a deployable compactor must be lightweight enough to call during inference, expressive enough to preserve context under constraint, and reusable across a trajectory. Existing compaction methods satisfy only part of this requirement: selection methods are lightweight but subset-bound, while synthesis methods are expressive but rely on per-context optimization. Here we introduce Still, a small per-layer Perceiver trained once against a frozen base model that produces compact keys and values in a single forward pass. On Qwen and Gemma models, Still occupies the favorable side of the speed--quality frontier across compression ratios from $8\times$ to $200\times$ and context lengths from $8$k to $128$k. On the long-context RULER grid, Still exceeds the strongest baseline by 8--22 points. The same compact cache also supports free-form summarization, preserving most of the full-context gain on HELMET and winning a pairwise LongBench summarization comparison against KV-Distill. Because compaction is a forward pass, Still can be applied iteratively, entering a long-horizon regime unavailable to per-context methods. We show that amortization makes long-context cache compaction tractable, and synthesis makes its compact state useful at extreme compression.

Large language models are increasingly used for social decision-making situations that require balancing cultural norms with personal preferences. For example, a user preferring honesty might ask whether to correct a coworker publicly when local norms favor indirect feedback. Yet existing research studies cultural alignment and personalization largely separately. We introduce PACT, the Personal-Preference and Cultural-Norm Trade-off framework, which evaluates whether models choose to follow a cultural norm or allow personal preferences. We find that LLMs vary in how rigidly they enforce cultural norms, with behavior shifted more by country context (7.8%) than age (1%) and gender (0.7%) and shifting non-uniformly after instruction tuning. Furthermore, our five-country human study on PACT shows that culture-following in humans is mainly driven by scenario country, with the lowest agreement when participants judge their own cultural contexts, showing within-culture pluralism. Finally, human-LLM alignment experiments show that models can match majority choices, but fail to capture response distributions and uncertainty (with best correlations reaching only 0.24). Together, these findings motivate alignment evaluations that go beyond majority to capture cultural pluralism and disagreement in social judgment.

LLMs-as-judges are the only way to evaluate safety at scale. Despite their importance, LLM-judges themselves are rarely evaluated beyond human agreement in simple, static benchmarks. We therefore investigate two under-explored but crucial properties of LLMs-as-judges: their susceptibility to relying on in context-information, and their steerability to differing safety definitions, which may not align with their internal safety priors. We evaluate the safety judging abilities of many generalist LLMs and safety-specific judges, and investigate the impact of task demonstrations, novel in-context information, and changing safety definitions. We find that while LLM-judges can learn from new information, they are broadly unlikely to adjust their evaluations if the context or safety definition contradicts their prior.

When a multimodal large language model answers a visual reasoning question correctly, is the prediction actually supported by the task-critical visual evidence? Correct answers can coexist with flawed reasoning, making accuracy alone an incomplete test of grounding. We introduce VisualFLIP, a paired benchmark with 1,374 images arranged as same-question perturbation pairs across cardinality, attribute, spatial, and logic tasks. Each pair keeps the question fixed but minimally changes the evidence so the gold answer deterministically flips. We evaluate 24 MLLMs with pair accuracy, which requires solving both sides of a pair, and Collapse Rate (CR), which measures how often a model that solves at least one side repeats the same non-empty answer for both images. Together, these metrics show that paired correctness and evidence dependence are related but distinct: capable models can still fail to update after task-critical visual changes, and collapse becomes more severe for some models when the edited image follows an earlier answer in a sequential setting. Further details are available on our project page: https://didizhu-judy.github.io/VisualFLIP/

Are tool-calling LLM agents equally safe throughout a conversation? We discover they are not: agents are most vulnerable at the very start of a session and become substantially safer after a few regular agentic tasks -- a phenomenon we term the cold-start safety gap. To study this systematically, we introduce Safety Over Depth for Agents (SODA), a benchmark that controls how many regular agentic tasks the agent completes before encountering a safety threat, supporting up to 20 preceding tasks. Evaluating 7 models from 4 families, safety improves by 9--52% as the number of preceding regular agentic tasks increases from zero to twenty. Representation analysis confirms that model hidden states gradually shift toward a safety-aligned region as more preceding tasks are present. By systematically studying which part of the preceding conversation matters most, we find that the regular agentic tasks themselves are the primary driver of safety, while the agent's own prior responses have less effect on safety but are essential for preserving later utility. This conclusion is further supported by evaluation on open-source safety benchmarks (AgentHarm, Agent Safety Bench) and utility benchmarks (BFCL, API-Bank), confirming that warming up the agent with regular agentic tasks before deployment makes it safer and preserves full capability. Based on these findings, we recommend a simple deployment strategy: having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests mitigates the cold-start safety gap. Our code is available at https://github.com/Trustworthy-ML-Lab/Agent-Cold-Start-Safety-Gap

Regulatory review of advanced nuclear reactor designs routinely spans more than three years and consumes hundreds of millions of dollars in combined regulator and applicant labor. We present the Regulatory Context Protocol (RCP), an Agent-to-Agent communication standard that replaces the formal human-to-human pipeline between regulators and applicants with a structured, auditable agentic channel, while preserving human oversight at safety-significant decision points. The protocol is calibrated against an analysis of 1,236 documents from U.S. Nuclear Regulatory Commission advanced reactor dockets and demonstrated with a working multi-agent pilot. Against an 89M USD, 42-month Reconstructed Baseline, RCP cuts costs by 50-77 percent (21M-44M USD) and timelines by 65 percent (15 months). Without a shared protocol, Standalone Agents reach only 54M-74M USD and 21 months. The residual cost-and-time gap is structural, not algorithmic: it traces to the inter-organizational pipeline that only an agent-to-agent standard can compress. The same bottleneck - formal multi-party review under strict auditability requirements - characterizes pharmaceutical approvals, environmental permitting, financial supervision, and aviation certification. The US regulatory paperwork burden carries a 426.5 billion USD annual opportunity cost; replicated broadly, the projected 50-77 percent reduction implies savings on the order of 210-330 billion USD per year - approaching 1 percent of US GDP.

Scientific machine learning is limited less by model size than by the data it is trained on. Observational data records what happened but not why; template synthetic data has a known generating process but only for the simulator's template, not the case a user faces. We argue a third option is now operationally feasible: instrumented data, in which every datum carries the mechanistic model that produced it, an explicit uncertainty over that model, and an executable family of counterfactuals. Verification-and-validation (V&V) instrumented image-to-simulation pipelines are one realisation: a sensor observation becomes a fully specified, solver-backed simulation with explicit, editable parameters and a propagated aleatoric/epistemic uncertainty. The substrate is case-specific, mechanistically supervised, and supports causal interventions through Pearl's do-operator. Near-term consequences for validation, auditing, and surrogate training span computational biology, climate, materials, fluid mechanics, and medical imaging; a longer-term, falsifiable implication concerns foundation models for scientific reasoning.

Recent vision-language models (VLMs) excel at multimodal understanding and reasoning, yet their fine-grained visual perception remains underexplored. A natural extension of ``How many r are there in Strawberry?'' asks: how small a visual pattern can a VLM reliably perceive? As such, we introduce FineSightBench, a new benchmark that systematically probes this limit by separating perception tasks (pixel-level recognition of letters, shapes, objects) from reasoning tasks (spatial reasoning, counting, ordering over small targets) across controlled scales of 4--48px. Through comprehensive experiments and detailed failure mode analysis on state-of-the-art models, we reveal a sharp dissociation: perception saturates around 12px, while reasoning remains limited even at larger scales, with persistent numeracy and sequence errors. These findings expose fundamental deficiencies in VLMs' fine-scale visual reasoning that demand more rigorous evaluation.

When a language model trains on its own verified outputs, does it acquire capability beyond its base, or merely get better at expressing capability the base already had? We make the question decidable with a teacher-free "constellation" -- a generator, a learned critic, and a free exact verifier -- on a FlashFill-style "trapdoor" DSL, where verified (problem, solution) pairs are cheap to synthesize, hard to invert, and free to check exactly. Everything runs on one 4-bit Qwen3-4B on a single 24 GB GPU, with no model in the loop larger than the base. We report three findings. (i) Critic-guided selection beats verifier-filtered best-of-$k$ by $+9.1$ pp ($6/6$ seeds), with the entire gain localized to tasks where candidates disagree on held-out inputs. (ii) Per-round STaR self-training raises the ceiling but never accelerates -- the gain tracks remaining headroom and decelerates across $K=4$ independent training trajectories. (iii) The domain has no clean zero-capability frontier, so the usual "$0\% \to$ climb $=$ emergence" test is invalid here. A measured pass@$K$ crossover settles the diagnosis: the trained model wins at the operating budget (pass@$8$) but the base overtakes it at a large budget (pass@$64$) on every trajectory, so self-training concentrates probability mass rather than expanding reach. This is amplification, not compounding. ($K=4$ is indicative, not yet a robust across-trajectory CI.)

Path-planning for autonomous vehicles in threat-laden environments is a fundamental challenge because the problem is nonlinear and nonconvex even in simplest scenarios. While traditional optimal control methods can be used to find ideal paths, the computational time is often too slow for real-time decision-making. To solve this challenge, we propose a method based on Deep Deterministic Policy Gradient (DDPG) and model the threat as possibly multiple circular 'no-go' zones. A mission is regarded as a failure if the vehicle enters this restricted zone at any time or does not reach a neighborhood of the destination. The DDPG agent is trained through trial and error in a simulated environment, learning a direct mapping from its current state (position and heading) to a series of feasible actions that guide the agent to safely reach its destination. The reword function has three parts: (a) an attractive field centered at the final destination, (b) some repulsive fields centered at the origins of circular obstacles, and (c) a penalty of control energy consumption (the magnitude of heading change) that indirectly in favor for straight path. The DDPG trains the agent using these incentives to find the largest possible set of starting points wherein a safe path to the destination is guaranteed. This provides critical information for mission planning, showing beforehand whether a task is achievable from a given starting point, assisting pre-mission planning activities. The approach is validated in simulation. A comparison between the DDPG method and a traditional optimal control (pseudo-spectral) method is carried out. The results show that the learning-based agent produces effective paths while being significantly faster, making it a better fit for real-time applications.

Large Language Models are transforming the support for clinical decision and their application in real scenarios. Yet, most benchmarks are conducted in English, and cross-lingual evaluation is needed to tackle the language gaps in global access. We introduce ClinicalBr, the first bilingual benchmark for clinical decision built from real Brazilian case reports. The corpus contains 2,892 cases drawn from 28 SciELO medical journals, spanning 18 specialties, and is structured as parallel Portuguese-English pairs. Each case supports four evaluation tasks: diagnosis retrieval, differential diagnosis, exam recommendation, and treatment planning. We evaluate four models: MedGemma-27B, Sabiá-4, DeepSeek-R1, and o3-mini, across both languages. The central finding is that the Portuguese-English performance gap is task-dependent, not general. In diagnosis retrieval, English yields a consistent advantage across all models, with +7.5-12.1 accuracy points. This advantage disappears in differential diagnosis, exam recommendation, and treatment planning, where confidence intervals cross zero for most models and Portuguese completeness scores are marginally higher. Brazilian-endemic conditions proved easier than the full corpus, not harder, indicating that tropical presentations are adequately represented in current pre-training. Exam recommendation was the hardest task across all models and both languages, with F1 scores below 0.10, well below the differential diagnosis ceiling of 0.20-0.27.

As language models improve and become increasingly deployed to solve a variety of tasks, trustworthiness becomes essential. Calibration is a good proxy for trust: well-calibrated confidence estimates help inform the risk versus reward tradeoff when trusting a specific model output. Unfortunately, even as models improve, they remain poorly calibrated, often biasing towards overconfidence. Additionally, calibration can be gamed: a policy that always predicts the base rate is perfectly calibrated, but completely uninformative. To resolve this, we develop a new metric, expected utility renormalized by the oracle (EURO), that balances calibration and informativeness. We also propose a general-purpose activation-based confidence, utility, and trust estimation protocol (ACUTE) to appropriately adjudicate uncertainty. The ACUTE protocol provides flexible, sample-efficient, and compute-efficient confidence estimators for 3 tasks including multiple choice question answering, tool-calling, and scientific document summarization across 6 models from 4 model families. ACUTE outperforms strong baselines on EURO, while maintaining low calibration error. Taken together, our work shows that equipping LLMs with the ACUTE protocol can improve calibration, utility, and trustworthiness in numerous settings.

Recently, the efficiency of Large Language Models (LLMs) deployment has become a critical concern in practical applications. While post-training quantization (PTQ) and structural pruning are established techniques for reducing memory footprint and inference latency, most existing PTQ approaches optimize quantization errors on a per-layer basis, overlooking how errors accumulate and propagate through the network, often resulting in suboptimal solutions. Traditional pipelines also tend to apply pruning and quantization in isolation or sequentially, further compounding sub-optimality. We introduce a novel end-to-end framework that addresses these limitations in two key ways. First, we propose a novel mixed-precision PTQ strategy that directly minimizes global error propagation across the entire model, rather than isolating layer-wise errors. Building on this, we develop a novel joint optimization approach that simultaneously learns structural pruning decisions and mixed-precision quantization policies within a unified search space. Extensive experiments show that, at ultra-low precisions (1-3 bits), our quantization method reduces WikiText perplexity by up to 21% compared to state-of-the-art (SoTA) weight-activation quantization baselines. Against leading weight-only quantization methods, it achieves up to 59% and 85% lower perplexity on WikiText and C4, respectively. Compared to the SoTA joint pruning-and-quantization techniques, our proposed method delivers superior perplexity and reasoning performance at ultra-low bits.

Navigation using a monocular camera is pivotal for autonomous operation on tiny aerial robots due to their perfect balance of versatility, cost and accuracy. In this paper, we introduce MinNav, a navigation stack based on optical flow and its uncertainty to fly through a scene with static and dynamic obstacles and unknown-shaped gaps without any prior knowledge of the scene components and/or their locations/ordering. We further improve success rate by using the activeness of the robot to move around in an exploratory way to find obstacles and navigate. We successfully evaluate and demonstrate the proposed approach in many real-world experiments in various environments with static and dynamic obstacles and unknown-shaped gaps with an overall success rate of 70%. To the best of our knowledge, this is the first solution to tackle all the aforementioned navigation cases without prior knowledge using a monocular camera. Our approach is on par in performance with depth based methods with factors of magnitude less computation required and can readily run onboard tiny aerial robots. The accompanying video, supplementary material, code and dataset can be found at https://pear.wpi.edu/research/minnav.html

Humanity is a mosaic of multifaceted talents and needs, and any truly intelligent AI must reflect that richness. Yet the LLMs used by all are built by the few -- a centralized market of monolithic AI models structurally ill-suited to capture the diversity of human knowledge, reasoning, and values. Here we introduce scaling participation, a new paradigm in which modular AI systems are built from the bottom up through the contributions of diverse stakeholders. Participants contribute small models trained on their own interests and priorities; these models then collaborate in modular frameworks as compositional AI systems. Participatory AI systems outperform monolithic LLMs by up to 15.4% across 15 tasks, such as reasoning and factuality, surpassing models larger than all contributed components combined. Further experiments show that participatory AI systems benefit from contributor diversity, substantially improve on each contributor's original priorities, and exhibit emergent capabilities that allow them to solve over 15% of problems where all individual models fail. Scaling participation provides a technical foundation for transitioning from the monolithic status quo toward an open, bottom-up, and collaborative AI future.

Large language models (LLMs) are widely used as judges for evaluating model outputs, but their high cost, latency, and opacity limit scalability. We introduce SLMJury, a framework for evaluating small language models (SLMs) as judges across two paradigms: closed-ended binary correctness and open-ended quality scoring. We benchmark 16 SLM judges (0.6B-14B parameters) from four model families across ten benchmarks: eight closed-ended tasks spanning mathematical, scientific, and general reasoning (N=64,824 judgments per configuration), plus SummEval and MT-Bench for summarization and conversational scoring. We formalize judging as a budget-conditioned function and study five dimensions. Four findings emerge. (1) The overthinking effect is domain-dependent: for most judges quick 10-token verdicts match or beat extended reasoning on mathematical judging (by 2-7% where they help), while reasoning wins on general tasks by up to 23%. (2) Domain generalization separates model families, with math-to-general accuracy gaps ranging from under 10% to nearly 40%. (3) Closed-ended and open-ended judging draw on different capabilities: the best binary judge (Phi-4) drops to rank 9 on MT-Bench, while reasoning-trained models invert this ordering. (4) Under the Reflect-Critique-Refine (RCR) debate protocol, multi-agent debate degrades accuracy across all tested configurations, whereas the top judges resist six adversarial personas with <=0.55% variance. Reliable automated evaluation does not require large proprietary models, yet no single SLM dominates. The leaderboard is available at https://anishh15.github.io/SLMJury/, and our framework code and pip package are publicly available at https://github.com/anishh15/SLMJury and https://pypi.org/project/slmjury/.

Reasoning language models deployed in agentic workflows must follow an instruction hierarchy: when instructions from different sources conflict, the model should obey the highest-privilege applicable instruction. Existing benchmarks largely measure this behavior end-to-end, asking whether the final response is compliant. However, a non-compliant response can arise from several distinct failures: the model may fail to identify the relevant instructions in context, fail to resolve conflicts among identified instructions, or correctly resolve the conflict in its reasoning while still producing a violating response. We introduce a white-box diagnostic framework that localizes instruction hierarchy failures into instruction identification, conflict resolution, and response realization, making failures more interpretable. We evaluate three reasoning models--Gemma-4-31B-IT, Qwen3.6-35B-A3B, and Claude Sonnet 4.6--on long-context adaptations of IHEval and IHChallenge, and find that the dominant failure mode varies across models, tasks, and context length. Building on the observation that models can often detect conflicts and output violations when explicitly prompted, we propose two training-free self-monitoring mechanisms: a parallel input monitor for low-latency conflict detection before generation, and a sequential output monitor for response-level review and repair. Across Gemma-4-31B-IT, Claude Sonnet 4.6, and GPT-5.3, the strongest monitor reduces rule-following non-compliance by 81-99%, with GPT-5.3 reductions of 86% under static attacks and 45% under adaptive attacks.

The rapid evolution of Large Language Models (LLMs) from passive assistants to autonomous, execution-capable agents has introduced critical operational risks. Most current evaluation frameworks neglect procedural compliance, leading to ''Machiavellian'' behaviors where agents strategically violate safety rules to maximize rewards - a direct manifestation of Goodhart's Law. To address this blind spot, we introduce MAC-Bench, a dynamic, adversarial benchmark designed to evaluate the procedural alignment of multi-agent systems under realistic pressure. We propose the SERV(Seed - Evolve - Refine - Verify) pipeline, an ``Agent-as-a-Benchmark'' paradigm that transforms unstructured legal texts into executable, contamination-free scenarios. By synthesizing holographic sandbox environments and injecting calibrated social-engineering pressure vectors, MAC-Bench forces agents into Pareto-optimal trade-offs between task success and regulatory adherence. We introduced novel metrics: the Compliance-Weighted Success Rate (CSR) and the Machiavellian Gap (MG), and conducted a comprehensive evaluation of state-of-the-art frontier models to reveal the pervasive trade-offs between success and compliance.

Multimodal reasoning requires a path that retains integrity over a wide range of constraints, from visual grounding to logic consistency. However, the current Process Reward Models focus on heuristically defined rewards that equally weigh these factors, which may lead to the concealment of individual dimension failures by the dominating factors, without guaranteeing the validity of the reasoning process in general.

Alzheimer's disease is a progressive neurodegenerative disorder, and its progression varies substantially across patients. Existing work aims to forecast patients' future cognitive state, with minimal focus on reconstructing the state from past visits. Furthermore, in current research, quantifying predictive uncertainty remains underexplored and relies on costly modalities such as MRI, PET, and CSF, limiting their deployment in resource-limited settings. In this research, our primary objectives are: First, bidirectional prediction of cognitive scores from irregular visits to present the complete disease trajectory. Second, to enable interpolation and extrapolation capabilities to assist clinicians in informed prognostic decision making, and third, to provide a well-calibrated uncertainty estimate for all predictions, and finally, to achieve the objectives using the modalities available during routine visits. We propose a unified framework, GNOVA: A GRU-Neural ODE Variational Autoencoder. The architecture combines a Gated Recurrent Unit encoder and a Neural ODE decoder within a variational autoencoder framework. In our work, we forecast the CDR-SB and MMSE Scores. The GRU encoder allows for any number of inputs at any time point. The Neural-ODE decoder performs continuous estimation, allowing interpolation and extrapolation at any desired time point. The Variational autoencoder allows for uncertainty estimation in predictions. We worked with 1,727 patients from the ADNI dataset over 10 years; the model achieved mean absolute errors of 1.35 and 2.28 for CDR-SB and MMSE scores, respectively, without requiring any neuroimaging or biomarker data. Feature-ablation studies revealed that age, BMI, and APOE4 status were strong predictors. The proposed framework enables the reconstruction of incomplete patient histories and the anticipation of future cognitive states.

Multi-agent LLM systems increasingly rely on communication protocols for coordination, yet their robustness under adversarial and structural constraints remains poorly understood. Building on prior work showing that cheap-talk channels enable cooperation in LLM coordination games, we investigate two vulnerability classes in a 4-player Stag Hunt across six model families and 720 trials. First, when Byzantine agents signal cooperation but defect, non-Byzantine agents detect the betrayal within one round yet fail to adapt collectively: a substantial fraction continue cooperating despite repeated exploitation, unable to recover coordination due to the game's unanimity payoff structure. Second, explicitly restricting communication topology collapses cooperation, while applying identical restrictions silently preserves near-perfect cooperation. This establishes that coordination failure stems from agents' meta-reasoning about hidden information, not information loss itself. We identify two stable behavioral archetypes that replicate across all model cohorts: Defection-Prone models that switch permanently after betrayal, and Cooperation-Persistent models that continue cooperating at significant individual cost. These findings reveal concrete security vulnerabilities: communication channels can be exploited as adversarial injection vectors, and disclosing network topology to agents can degrade coordination even without any adversary present.

Data stream mining is fundamentally challenged by concept drift, where distributional changes can degrade model performance. Despite the proliferation of drift detection methods, progress in the field is hindered by inconsistent evaluation practices: studies rely on oversimplified synthetic data generators, adopt incompatible metrics, and lack transparency in hyperparameter selection, making fair comparisons difficult. We address this gap with a novel benchmarking framework comprising three contributions: (1) a drift simulation method that injects controlled distributional changes into real-world datasets via Monte Carlo trials, enabling supervised evaluation while preserving real-world data complexity; (2) an evaluation protocol for drift detection with timing-aware criteria, including the derivation of new metrics (e.g., F1 detection score, normalized detection time) that are comparable across streams; and (3) we advocate for a leave-one-dataset-out hyperparameter optimization protocol for drift detection methods that promotes configuration robustness across heterogeneous stream dynamics. We benchmark 14 widely used drift detection methods on 7 realworld datasets across 4 drift types (class prior, label swap, feature permutation, feature filtering), each under both abrupt and gradual transitions. Our experimental results provide insights into the strengths and weaknesses of current drift detection approaches while establishing baseline performance metrics for future research in this area. All code and experiments are publicly available.

Retrieval-Augmented Generation (RAG) is widely used to improve the factual reliability of large language models (LLMs) by grounding answers in retrieved evidence. In misinformation-rich environments, however, retrieved content may include plausible but incorrect information, raising concerns about the reliability of RAG-based information access systems. In this work, we propose an evaluation protocol to systematically test how the RAG system handles conflicts between parametric knowledge and evidence retrieved from context with varying amounts of misleading information. We target correct answers to factoid questions that the model responds to correctly, even when there is no retrieval, and use this to test the system with clean, poisoned, and mixed evidence. The proposed analytical framework combines parametric override and confidence metrics to assess when and how misleading information affects the generation process of LLMs. This study aims to provide insights into the robustness of RAG systems in information disorder scenarios.

Dominant web data curation pipelines for pretraining collapse document quality into a single composite score, systematically missing high-value content along dimensions the scorer underweights. We present a taxonomy-driven framework that recovers this value by filtering along semantically meaningful dimensions that composite scores fail to capture. First, building on the ESSENTIAL-WEB taxonomy, we introduce two novel dimensions: timeliness and cultural specificity, both of which show low pairwise NMI with existing ones. We annotate 14M documents using Qwen2.5 32B and distill into a lightweight 0.5B model. To enable rapid corpus-wide annotation, we additionally train a 73M multi-task MLP on E5 embeddings, achieving 50x inference throughput. Second, to navigate the combinatorial explosion of filter configurations, we introduce a compute-efficient two-pass framework: Pass 1 identifies the strongest dimension signals at small scale; Pass 2 constructs and evaluates conjunctive and disjunctive compound filters from the top performers - identifying high-performing configurations at a fraction of full scaling-law cost. Applying the selected filters to deprioritized web data, taxonomy-filtered subsets outperform their unfiltered baselines and even surpass the highest-quality tier. On mid-tier data, our best filter improves over its unfiltered baseline by 12.1% on reasoning, 9.5% on coding, and 2.0% on knowledge benchmarks, exceeding unfiltered top-tier data by 6.7% on reasoning and 13.7% on coding. Furthermore, filtered data from two tiers below the typical production threshold improves by 22.3% on reasoning and 19.5% on coding over its unfiltered baseline, surpassing top-tier data on coding benchmarks. These results establish that vast latent value remains locked in deprioritized web data, and that multi-dimensional taxonomy filtering is a principled, compute-efficient key to unlocking it.

Recent breakthroughs in self-supervised learning (SSL), such as the Latent-Euclidean Joint-Embedding Predictive Architecture (LeJEPA), alongside successes in integrating visual encoders with language models, have driven the demand for adaptable, high-capacity vision encoders in Computed Tomography (CT). In this work, we explore 2D slice-based architectures as a flexible alternative to native 3D models for processing volumetric CT data. Using the CT-RATE dataset, we trained DALE-CT (Depth-Aware Latent-Euclidean Computed Tomography), a 2D model family built entirely from scratch using LeJEPA, and compared its performance against a continually pre-trained DINOv2 baseline. To enhance representation quality, we developed a novel 3D depth-aware pre-training strategy anchored by dense auxiliary supervision from both automated anatomical masks and human-annotated abnormalities. Under linear probe evaluation with Multiple Instance Learning (MIL) for multi-abnormality detection, the frozen backbone of this dual-supervised model (DALE-CT-2S) achieves a Macro AUROC of 0.833. This performance demonstrates near-parity with state-of-the-art 3D vision-language models, achieved entirely from scratch with significantly less data and no textual supervision. To ensure reproducibility, all training code, evaluation scripts, and model weights have been made publicly available.

Self-supervised methods that learn representations and predict dynamics fully in the latent space, such as JEPA, have been shown to confuse slowly varying noise with the dynamical signals they aim to capture. Specifically, when noise features remain approximately constant within each trajectory, contrastive predictive objectives preferentially encode these features instead of the true latent variables governing the system. The learned representation then becomes dominated by trajectory-specific noise, so downstream performance degrades with noise strength and does not improve even as the number and duration of training trajectories increase. We argue that this failure is a property of the objective itself, shared by a long line of contrastive predictive objectives that sample negatives across trajectories. To illustrate this generality, we study the failure mode and its remedy in two settings: a standard SimCLR-style JEPA on a synthetic moving-dot dataset, and DySIB, a recently introduced method designed for extracting physically interpretable representations of dynamics, on movies of a rigid-body pendulum. When negatives are instead sampled within a single trajectory, the slow noise can no longer distinguish frames within that trajectory, removing the predictive shortcut. Training one encoder simultaneously on many such trajectories then forces it to encode the variables relevant for the dynamics, with longer trajectories yielding better representations even for strong slow noise. Our results point toward principles for designing contrastive predictive objectives in dynamical representation learning, especially for physical systems with noisy experimental observations.

Accurate distance estimation for small drones in long-range imagery is important for tracking and situational awareness, yet remains challenging due to extreme target scale variation, background clutter, and noisy visual cues. This paper studies monocular drone distance estimation using image crops together with bounding-box geometry, a practical setting in which a detector provides a candidate drone region and the model predicts range from appearance and box-derived features. We evaluate a Droneranger-style baseline, and introduce a new DroneDAR (Drone Detection And Ranging) model that combines a convolutional backbone with explicit bounding-box cues through a lightweight gating mechanism. Experiments analyze how backbone capacity, crop resolution, and regression loss functions affect performance across distance regimes. We further examine common failure modes at long distances, including sensitivity to bounding-box noise and reduced texture detail in the crop. The results provide guidance for designing and training range estimators that remain robust under real-world long-range conditions and highlight directions for improving reliability when drones occupy only a few pixels.

It is well established that ReLU networks define continuous piecewise-linear functions, and that their linear regions are polyhedra in the input space. These regions form a complex that fully partitions the input space. The way these regions fit together is fundamental to the behavior of the network, as nonlinearities occur only at the boundaries where these regions connect. However, relatively little is known about the geometry of these complexes beyond bounds on the total number of regions, and calculating the complex exactly is intractable for most networks. In this work, we prove new theoretical results about these complexes that hold for all fully-connected ReLU networks, specifically about their connectivity graphs in which nodes correspond to regions and edges exist between each pair of regions connected by a face. We find that the average degree of this graph is upper bounded by twice the input dimension regardless of the width and depth of the network, and that the diameter of this graph has an upper bound that does not depend on input dimension, despite the number of regions increasing exponentially with input dimension. We corroborate our findings through experiments with networks trained on both synthetic and real-world data, which provide additional insight into the geometry of ReLU networks. Code to reproduce our results can be found at https://github.com/bl-ake/ICLR-2026.

Large Language Models are typically benchmarked by evaluating every model on every test query. For practitioners seeking the best model to deploy, this is often wasteful: if a model clearly performs worse than others, there is no need to precisely estimate its performance. Best-arm identification algorithms can be naturally applied to drastically reduce costs by adaptively allocating evaluation budget. Further, language models often respond similarly to the same prompt-a property previous work has tried to leverage with mixed success. We propose Synchronized Successive Rejects (SySRs), augmenting the classical Successive Rejects algorithm with paired comparisons. Unlike prior attempts to leverage model similarity in best-model identification, our approach is hyperparameter-free and enjoys performance guarantees that improve with the degree of similarity between evaluated models. Empirically, our method outperforms all baselines in terms of average error rate across 15 standard benchmarks, and in terms of worst-case budget for reliably identifying the best model.